ChuckMcSneed/Gembo-v1-70b
ChuckMcSneed/Gembo-v1-70b is a 69 billion parameter experimental merge model developed by ChuckMcSneed, primarily designed for roleplay (RP) and general assistant tasks. This model, built upon a base of Sao10K/Euryale-1.3-L2-70B and multiple Llama-2-70B QLoRAs, exhibits strong creativity and moderate intelligence. It achieves a 70.51 average on the Open LLM Leaderboard, with notable scores in MMLU (70.85) and ARC (71.25), making it suitable for creative text generation and assistant applications.
Loading preview...
Model Overview
ChuckMcSneed/Gembo-v1-70b is a 69 billion parameter experimental merge model, primarily developed by ChuckMcSneed for roleplay (RP) and general assistant use cases. It was created by merging Sao10K/Euryale-1.3-L2-70B with several Llama-2-70B QLoRAs and then SLERP-merged with cognitivecomputations/dolphin-2.2-70b and ChuckMcSneed/SMaxxxer-v1-70b. The model is noted for its creativity and moderate intelligence, offering a balance between being lightly censored and capable of nuanced responses.
Key Capabilities & Characteristics
- Roleplay Optimization: Specifically designed with RP in mind, incorporating multiple LORAs focused on creativity.
- Assistant Functionality: Capable of serving as a general assistant, demonstrated by its performance in benchmarks.
- Creative & Verbose: Exhibits high creativity and can be quite verbose, though it may sometimes stop abruptly.
- Balanced Censorship: Described as "lightly censored" and "fairly neutral," avoiding extremes of uncensored or overly censored behavior.
- Robust Merging: Built from a complex merge of various models, including Guanaco, Saiga2, Airoboros, Kimiko, LimaRP, Wizard, and TonyGPT LORAs, indicating a broad training base.
Performance Highlights
- Open LLM Leaderboard: Achieves an average score of 70.51, with specific scores including 71.25 on ARC, 86.98 on HellaSwag, 70.85 on MMLU, and 63.25 on TruthfulQA. This indicates strong reasoning and general knowledge capabilities.
- WolframRavenwolf Benchmark: Successfully answered 18/18 multiple-choice questions correctly, demonstrating its utility for real-world assistant tasks, though it may not strictly follow single-letter response instructions.
- UGI Leaderboard: Scores 36.38, which is around average, suggesting a balanced approach to uncensored content.
Recommended Usage
- Prompt Format: Uses Alpaca format, but is flexible due to its merged nature.
- Settings: For optimal performance, especially in RP, higher temperature (e.g., 5) and repetition penalty (e.g., 1.15) are recommended, along with a MinP of 0.10.